Public health programs operate in a state of perpetual scarcity. Whether it is a clinic in Tanzania trying to retain HIV patients, a maternal health program in India sending "nudge" calls to new mothers, or a hospital in the US preventing pressure injuries, the problem is identical: the number of people who could benefit from an intervention far exceeds the staff, time, or budget available to provide it.
Traditionally, these resources are allocated using simple heuristics—such as calling every patient or focusing only on the most recently seen. However, a growing body of research suggests that we can significantly improve health outcomes by treating resource allocation as a predictive optimization problem.
From Risk to Impactability
The first step in optimizing interventions is risk stratification: identifying the small subset of a population that accounts for the majority of adverse events. Warren et al. demonstrated this in the context of opioid use, showing that the highest-risk 10% of patients account for 35% of overdoses. Similarly, Perfalk et al. and Bishara et al. have shown that machine learning models trained on electronic health records (EHR) can identify patients at high risk for psychiatric readmission or postoperative delirium far more accurately than traditional clinical scores.
However, knowing who is at risk is not enough. A patient might be at high risk but unlikely to change their behavior regardless of the intervention. Conversely, a low-risk patient might be self-motivated and succeed without help. Torkamani et al. introduce the concept of impactability: prioritizing individuals who are most likely to respond to an outreach effort. By targeting the "impactable" middle—those who need a nudge to act but are capable of doing so—outreach programs can achieve a nearly 300% improvement in efficiency compared to traditional methods.
Managing Dynamic Risk with Restless Bandits
In many health settings, risk is not static; it evolves over time. A mother might be highly engaged with a health program one week and drop out the next. To manage these "restless" dynamics, researchers have turned to the Restless Multi-Armed Bandit (RMAB) framework.
In an RMAB system, each patient is modeled as a "bandit arm" that can be in different states (e.g., engaged or disengaged). The state changes over time according to different probabilities depending on whether the patient receives an intervention (a "pull" of the arm). Because only a few patients can be intervened upon at once, the system must decide which "arms" to pull to maximize the long-term health of the entire population.
Biswas et al. proposed the Whittle Index based Q-Learning (WIQL), which allows health systems to learn these transition probabilities on the fly. This is crucial because, as Verma et al. describe in the deployment of the SAHELI system, real-world patient behavior is often unpredictable. By using RMABs to schedule health worker calls, SAHELI reduced engagement drops by 30.5% for over 100,000 mothers in India.
Scaling to Millions
While RMABs and other optimization frameworks like Integer Linear Programming (used by Kehinde et al. for vaccination uptake in Nigeria) are powerful, they are computationally expensive. Solving these problems for hundreds of thousands of patients in real-time is a massive technical hurdle.
To address this, Killian et al. developed the GROUPS framework, which clusters similar individuals together to reduce the complexity of the optimization. This allows robust decision-making to scale by over 1,000x, making it possible to manage large-scale programs while remaining resilient to the noise and uncertainty inherent in public health data.
Beyond Individual Outreach
The logic of predictive prioritization extends beyond individual phone calls to broader systemic choices:
- Screening Intervals: Brentnall et al. showed that by extending screening intervals for low-risk individuals and shortening them for high-risk ones, cancer incidence can be reduced without increasing the total number of mammograms.
- Social Networks: Yadav et al. developed the HEALER agent, which selects "peer leaders" among homeless youth to maximize the spread of HIV prevention information through social networks, accounting for the fact that the network structure itself is often uncertain.
- Disease Surges: During high-volume events like the COVID-19 pandemic, Conderino et al. argue that systems must pivot from universal tracing to targeted interventions in high-risk settings to prevent the total collapse of outreach capacity.
By shifting from reactive care to predictive prioritization, these researchers demonstrate that we can do more with less—not by working harder, but by using data to ensure that every limited intervention is directed where it will have the greatest impact.
Key Foundational Work
Modeling Transmission Dynamics and Control of Vector-Borne Neglected Tropical Diseases (2010) — contributed 1 concept elements to this direction.
Modeling and Mining Spatiotemporal Patterns of Infection Risk from Heterogeneous Data for Active Surveillance Planning (2014) — contributed 1 concept elements to this direction.